🎙️ Gemini 2.5 Pro Audio Transcription & Diarization

This project showcases how to use Gemini 2.5 Pro for transcribing multi-speaker audio, applying audio diarization (identifying "who said what"), and working with long-form audio files effectively.

---

🔍 Overview

With Gemini 2.5 Pro, you can:

• Upload and transcribe long-form audio (up to 9.5 hours in total per prompt).
• Perform audio diarization to segment and label speakers automatically.
• Handle common audio formats like .wav, .mp3, .aac, etc.
• Work within token and resolution limits using smart chunking.
• Extract meaningful text output with speaker attribution.

---

🧠 How Audio Diarization Works

Input Audio (Multi-speaker)
       |
1. Feature Extraction — Extract MFCCs, spectral features
       ↓
2. Voice Activity Detection (VAD) — Detect speech vs. non-speech
       ↓
3. Speaker Segmentation & Clustering — Identify unique speakers
       ↓
Output: Speaker-Labeled Timeline

Example:

[00:00–00:15] Speaker A: Hello and welcome...
[00:15–00:45] Speaker B: Thanks! I'm excited...

---

📁 Supported Audio Formats

Gemini supports:

• WAV - audio/wav
• MP3 - audio/mp3
• AIFF - audio/aiff
• AAC - audio/aac
• OGG - audio/ogg
• FLAC - audio/flac

---

🧩 Tokenization & Audio Limits

• 1 second of audio ≈ 32 tokens
• 1 minute of audio ≈ 1,920 tokens
• Max input tokens = 1,000,000
• Max output tokens = 64,000

Audio Duration	Input Tokens (approx.)	Output Tokens
15 minutes	28,800 tokens	~8K tokens
1 hour	115,200 tokens	~32K tokens
2 hours	230,400 tokens	~64K tokens

🔁 To process more than 2 hours of audio, chunk the file and use prompt windowing (e.g., from X min to Y min).

---

💾 Uploading Audio to Gemini

You can provide audio in two ways:

• Pre-upload via API or file handler.
• Inline data in the request body.

---

NB: I removed the lines

import IPython.display as ipd  
ipd.Audio("/content/HS4830417304.mp3", autoplay=True)  

so that the file can be uploaded to Git—it was beyond the 25MB limit.
Feel free to add those lines at the end of the notebook if needed.

💵 Pricing (as of Gemini 2.5 Pro)

Modality	Price / 1M Tokens
Input ≤ 200K	$1.25
Input > 200K	$2.50
Output ≤ 200K	$10.00
Output > 200K	$15.00

---

🧪 Additional Notes

• Audio is downsampled to 16 kbps mono. Stereo channels are merged.
• Gemini understands non-speech audio, but only returns responses for English-language speech.
• You can label speakers manually if diarization doesn’t identify names correctly.

---

🧰 Example Use Case

prompt = """
Please transcribe the audio and indicate speaker turns.

From minute 5:00 to minute 10:00:
Speaker A is Alice
Speaker B is Bob
"""

---

📈 Output Example

[Alice]: Welcome to the show!
[Bob]: Thanks, it’s great to be here.
...

---

🚀 How to Use (Coming Soon)

• Python example with Gemini API
• Audio file chunking helper
• Diarization label enhancer

---

📦 File Summary

• Gemini_2_5_Pro_Podcast_Audio_transcription_boo.ipynb – Jupyter notebook showing audio processing
• Images/screenshots from Sam Witteveen’s tutorial

---

👤 Credits

• Gemini AI by Google DeepMind
• Visuals inspired by Sam Witteveen
• README by Emmanuel Kasigazi

---